Prediction of cholesterol level in patients with myocardial infarction based on medical data mining methods
Cemil Colak1,*, Mehmet C. Colak2, Necip Ermis3, Nevzat Erdil2, Ramazan Ozdemir3
1Dept. of Biostatistics and Medical Informatics, Inonu University, Faculty of Medicine, Malatya, Turkey
2Dept. of Cardiovascular Surgery,Inonu University, Faculty of Medicine, Malatya, Turkey
3Dept. of Cardiology, Inonu University, Faculty of Medicine, Malatya, Turkey
*Corresponding author: cemilcolak@yahoo.com
Abstract
Myocardial infarction (MI) is a significant reason for death and disability over the world and might be the first sign of coronary artery disease. The current study was carried out to predict the cholesterol level in patients with MI using data mining methods, artificial neural networks (ANNs) and support vector machine (SVM) models. The data of 596 patients, who had been diagnosed with segment elevation MI were analysed in the present study. The retrospective dataset including gender, age, weight, height, pulse, glucose, creatinine, triglyceride, high-density lipoprotein, and low-density lipoprotein was used for predicting the cholesterol level. Correlation based feature selection was applied.
Multilayer perceptron (MLP) ANNs and SVM with radial basis function kernel were used for the prediction based on the selected predictors. The performance of the ANNs and SVM models was evaluated on the basis of correlation coefficient and mean absolute error. The estimated correlation coefficients observed and predicted values were 0.94 for ANNs and 0.88 for SVM in training dataset (n=376), and 0.95 for ANNs and 0.90 for SVM in testing dataset (n=160), respectively. ANNs and SVM models yielded mean absolute error of 7.37 and 14.18 in training dataset, and 7.87 and 14.71 in testing dataset, consecutively. The results of the performance evaluation showed that MLP ANNs performed better for the prediction of cholesterol level in patients with MI in comparison to SVM. The proposed MLP ANNs model might be employed for predicting the level of cholesterol for MI patients in clinical decision support process.
Keywords: Artificial neural networks (ANNs); cholesterol level; medical data mining; myocardial infarction (MI);
support vector machine (SVM).
1. Introduction
Myocardial infarction (MI) can be perceived by clinical peculiarities and is a real reason for death and handicap around the world. MI may be one of the first indications of coronary artery disease (CAD). MI may have major mental and legitimate ramifications for the society and is an indicator of one of the health problems over the world (Thygesen et al., 2012).Worldwide, cardiovascular sickness is assessed to be the main reason for death.
Powerful prevention needs a worldwide strategy focused on the importance of risk factors for cardiovascular sickness in diverse geographic districts (Yusuf et al., 2004).
The reasons for CAD are multifactorial. Some of these variables are modified and are associated with ways of life; for example, tobacco smoking, absence of physical
movement, and dietary propensities. Other variables are non-modifiable, such as age and male gender. Studies discovered an immediate relationship between levels of low-density lipoprotein (LDL) cholesterol and the rate of new-onset CAD in males and females, who have no CAD at first (Investigators, 1992; Stamler et al., 1986; Wilson et al., 1998). The same relation holds for repetitive coronary occasions in individuals with situated CAD.
Among the supervised machine learning methods, artificial neural networks (ANNs) are computer based programs, and accumulate their knowledge from input- output relationships in datasets (Colak et al., 2008). Support vector machine (SVM) is one of the supervised machine learning methods used widely in pattern recognition and classification problems and performs a classification by building a multidimensional hyperplane that ideally segregates between two classes (Yu et al., 2010).
Of the studies related with estimation of the probability of MI, a study used logistic regression and ANNs models based on patient clinical history variables, and reported that both models can predict successfully the likelihood of myocardial infarction according to some factors alone (Wang et al., 2001). Another study assessed an improvement achieved by ensemble-based methods, bootstrap aggregation of regression trees, random forests, and boosted regression trees in patients with either acute MI or congestive heart failure. The study reported that ensemble methods of data mining boosted the prediction performance of regression trees.
To our knowledge, no study has been reported on the prediction of cholesterol level in MI patients using medical data mining methods. Consequently, the current study attempted to predict the cholesterol level in patients with MI using medical data mining methods, ANNs and SVM models.
2. Materials and Methods 2.1. Dataset
The studied data consisted of a sample without replacement from the database of Cardiology Department of Turgut Ozal Medical Center, Malatya, Turkey, between 2010 and 2013. The data of 596 patients, who had been diagnosed with segment elevation MI in pursuant of second universal definition of myocardial infarction guideline (Thygesen et al., 2007a; Thygesen et al., 2007b) were analysed in the present study. The retrospective dataset included cholesterol level, gender, age, weight, height, pulse, glucose, creatinine, triglyceride, high-density lipoprotein (HDL), and LDL.
Table 1 defines the details of the target and input attributes.
Table 1. The details of the target and input attributes
Attribute Attribute type Role Cholesterol level Numerical Target Gender (female/male) Categorical Input
Age (years) Numerical Input
Weight (kg) Numerical Input
Height (m) Numerical Input
Pulse (beats per minute) Numerical Input Glucose (mg/dL) Numerical Input Creatinine (mg/dL) Numerical Input Triglyceride (mg/dL) Numerical Input
HDL (mg/dL) Numerical Input
LDL (mg/dL) Numerical Input
2.2. Knowledge discovery in databases process
The knowledge discovery in databases (KDD) process is given in Figure 1.
Fig. 1. An overview of the data mining step and additional steps in the KDD Process (Fayyad et al., 1996)
According to Figure 1, KDD process contains five steps:
Data selection: Selecting data related to the analysis task from the database.
Data pre-processing: Removing outliers, extreme values, noise and inconsistent data.
Data transformation: Transforming data into convenient structures to implement data mining.
Data mining: Choosing data mining algorithm(s) being suitable to pattern in the data; extracting data patterns.
Evaluation and interpretation: Identifying the most suitable model(s) to obtain the targeted knowledge (Silwattananusarn & Tuamsuk, 2012).
2.3. Power analysis and software
The power analysis calculated minimum 265 subjects with the supposed cholesterol difference of 5, assumed standard deviation of 25, type I error (α) of 0.05 and type II error (β) of 0.10.For analysing and modelling the data, IBM SPSS Modeler Professional 16.0 for Windows was employed.
3. Results
The current study initially included 298 (50.0%) male and 298 (50.0%) female MI patients, 596 in total. Mean age of the patients was 68.3±12.3 y. The KDD is explained in the following steps.
3.1. Data Selection
The target was cholesterol level (numerical target attribute), and the predictors were gender, age, weight, height, pulse,
glucose, creatinine, triglyceride, high-density lipoprotein (HDL) and low-density lipoprotein (LDL).
3.2. Data pre-processing
Multivariate outliers and extreme values in the data were detected using T2 test based on the Mahalanobis distance.
This method uses a hypothesis testing based on the T2 probability levels to test multiple extreme values. The determined 60 inconsistent instances were discarded, and further analyses were performed on the remaining instances (n=536).
3.3. Data transformation and reduction
Feature selection based on correlation was carried out for reducing the predictors. The selected predictors were LDL, triglyceride, HDL, age and gender, respectively. The variables of weight, height, pulse, glucose and creatinine were not selected as a result of feature selection. The chosen numerical attributes of LDL, triglyceride, HDL and age were transformed to standard units (Mean=0, Standard Deviation= 1) called Z-transformation.
3.4. Data mining
Multilayer perceptron (MLP) ANNs and SVM with radial basis function kernel were used for the prediction of cholesterol level based on the selected predictors of LDL, triglyceride, HDL, age and gender. The applied ANNs structure was displayed in the Figure 2. MLP is one of the most popular neural network architectures and is a supervised network, owing to the fact that it calls for a desired output to learn. MLP includes an input layer with neurons (input variables), an output layer with neurons (target variable), and one or more hidden layers containing neurons to discover the nonlinearity in the data (Hongfei et al., 2013; Süt & Çelik, 2012). The MLP ANNs included a hidden layer with 5 neuron, and activation functions for the hidden output layers were hyperbolic tangent and identity, respectively. SVM is a supervised machine learning technique in using classification and regression routines (Moses, 2015; Shin et al., 2014), and showed a good performance in solving medical and biological classification and prediction problems (Arslan et al., 2016; Colak et al., 2015; Zhou et al., 2014). Among the kernel functions, RBF for SVM was selected in the current study. As a result of the implementation of grid search algorithm, the regularization parameter (C), regression precision (epsilon) and RBF gamma were determined as 10, 0.1 and 0.1, respectively.
Fig. 2. The structure of applied ANNs model (*: transformed variable)
Relative predictor importance for the chosen variables is presented in Table 2 on the basis of the results of ANNs and SVM models; while relative predictor importance for ANNs model in descending order was LDL, HDL, triglyceride, age and gender. The importance arrangement for SVM model in descending order was LDL, HDL, age, gender and triglyceride.
Table 2. Relative predictor importance for the selected variables
Relative predictor importance
Variables ANNs SVM
LDL* 0.55 0.30
HDL* 0.22 0.24
Triglyceride* 0.18 0.11
Age* 0.03 0.19
Gender 0.02 0.16
*: transformed to standard units (Mean=0, SD=1)
3.5. Evaluation and interpretation
Using holdout technique, the dataset was divided into two sets: 70% of the dataset (n=376) for training the models, and 30% of the dataset (n=160) for testing the models. The performance of the ANNs and SVM models was assessed on the basis of correlation coefficient and mean absolute error. Table 3 tabulates the details of the model evaluation.
The estimated correlation coefficients of observed and predicted values were 0.94 for ANNs and 0.88 for SVM in training dataset (n=376), and 0.95 for ANNs and 0.90 for SVM in testing dataset (n=160), respectively. ANNs and SVM models yielded mean absolute error of 7.37 and 14.18 in training dataset, and 7.87 and 14.71in testing dataset, consecutively.
Table 3. The details of the model evaluation
Training (n=376) Testing (n=160)
Model correlation coefficient
mean absolute
error
correlation coefficient
mean absolute
error
ANNs 0.94 7.37 0.95 7.87
SVM 0.88 14.18 0.90 14.71
4. Conclusions
The current study attempted to predict the cholesterol level in patients with MI using medical data mining methods, ANNs and SVM models. To achieve this objective, we used the knowledge discovery process for extracting knowledge from data. In the first and second steps, the data related to cholesterol level were selected, and since there were outliers and extreme values, the inconsistent data were removed to increase the prediction performance of cholesterol level. In the third step, the data were transformed to convenient structures to implement data mining, and a subset of relevant features was selected in the model construction. As for the fourth step, we applied two medical data mining algorithms, ANNs and SVM for extracting data patterns. Finally, when the values of correlation coefficient and mean absolute error were evaluated and interpreted, ANNs yielded the best performance as compared with SVM in training and testing datasets.
Coefficient of determination (R2) values were calculated as 0.902 for ANNs and 0.810 for SVM models. This finding also demonstrated that ANNs produced higher R2 value in the prediction of the cholesterol level in patients with MI, when compared with SVM model. As for the selected predictors, both models used 5 features selected out of 10 features for optimal prediction of cholesterol.
As a consequence, the results of the performance evaluation showed that MLP ANNs performed better for the prediction of cholesterol level in patients with MI in comparison to SVM. The proposed MLP ANNs model might be employed for predicting the level of cholesterol for MI patients in clinical decision support process.
References
Arslan, A.K., Colak, C. & Sarihan, M.E. (2016). Different medical data mining approaches based prediction of ischemic stroke. Computer Methods and Programs in Biomedicine, 130:87-92.
Colak, C., Karaman, E. & Turtay, M.G. (2015). Application of knowledge discovery process on the prediction of stroke. Computer Methods and Programs in Biomedicine, 119(3):181-185.
Colak, M.C., Colak, C., Kocatürk, H., Sağiroğlu, S. & Barutçu,
I. (2008). Predicting coronary artery disease using different artificial neural network models. Anadolu kardiyoloji dergisi: AKD= the Anatolian Journal of Cardiology, 8(4):249-254.
Fayyad, U., Piatetsky-Shapiro, G. & Smyth, P. (1996). From data mining to knowledge discovery in databases. AI Magazine, 17(3):37.
Hongfei, W., Yunyan, Z., Fei, Y. & Hui, L. (2013). Evaluation of an artificial neural network to ascertain why there is a high incidence of hepatitis B in the Chinese population after vaccination. Computers in Biology and Medicine, 43(9):1167-1170.
Investigators, L.R.C. (1992). The lipid research clinics coronary primary prevention trial: results of 6 years of post-trial follow-up.
Archives of Internal Medicine, 152(7):1399.
Moses, D. (2015). A survey of data mining algorithms used in cardiovascular disease diagnosis from multi-lead ECG data. Kuwait Journal of Science, 42(2):206-235.
Shin, J., Park, H., Cho, S., Nam, H. & Lee, K.J. (2014). A correction method using a support vector machine to minimize hematocrit interference in blood glucose measurements. Computers in Biology and Medicine, 52(0):111-118.
Silwattananusarn, T. & Tuamsuk, K. (2012). Data mining and its applications for knowledge management: A literature review from 2007 to 2012. International Journal of Data Mining & Knowledge Management Process, 2(5):13-24.
Stamler, J., Wentworth, D. & Neaton, J.D. (1986). Is relationship between serum cholesterol and risk of premature death from coronary heart disease continuous and graded?: Findings in 356 222 primary screenees of the multiple risk factor intervention trial (MRFIT). Jama, 256(20):2823-2828.
Süt, N. & Çelik, Y. (2012). Prediction of mortality in stroke patients using multilayer perceptron neural networks. Turkish Journal of Medical Sciences, 42(5):886-893.
Thygesen, K., Alpert, J.S., Jaffe, A.S., White, H.D., Simoons, M.L., et al.(2012). Third universal definition of myocardial infarction. Journal of the American College of Cardiology, 60(16):1581-1598.
Thygesen, K., Alpert, J.S. & White, H.D. (2007a). Universal definition of myocardial infarction. European Heart Journal, 28(20):2525-2538.
Thygesen, K., Alpert, J.S., White, H.D., Jaffe, A.S., Apple, F.S., et al. (2007b). Universal definition of myocardial infarction. Circulation, 116(22):2634-2653.
Wang, S.J., Ohno-Machado, L., Fraser, H.S.F. & Kennedy, R.L.
(2001). Using patient-reportable clinical history factors to predict myocardial infarction. Computers in Biology and Medicine, 31(1):1-13.
Wilson, P.W., D’Agostino, R.B., Levy, D., Belanger, A.M., Silbershatz, H., et al. (1998). Prediction of coronary heart disease using risk factor categories. Circulation, 97(18):1837-1847.
Yu, W., Liu, T., Valdez, R., Gwinn, M. & Khoury, M.J. (2010).
Application of support vector machine modeling for prediction of common diseases: the case of diabetes and pre-diabetes. BMC Medical Informatics and Decision Making, 10(1):16.
Yusuf, S., Hawken, S., Ôunpuu, S., Dans, T., Avezum, A., et al.
(2004). Effect of potentially modifiable risk factors associated with myocardial infarction in 52 countries (the INTERHEART study): case- control study. The Lancet, 364(9438):937-952.
Zhou, S., Li, G.B., Huang, L.Y., Xie, H.Z., Zhao, Y.L., et al. (2014).
A prediction model of drug-induced ototoxicity developed by an optimal support vector machine (SVM) method. Computers in Biology and Medicine, 51:122-127.
Submitted : 29/12/2014 Revised : 15/02/2016 Accepted : 17/02/2016
VO U √ ”U √ vK VKI « WKC ¡UA « s Êu UF s c « v{d*« w Âb « w ‰Ëd O uJ « Èu 0 R M « WO D « U UO « rOGK
dO œ“Ë√ ÊUC —
3¨q œ—≈ «eHO
2¨fO —≈ VO O
3¨„ôu e MO XLN
2¨„ôu qOLO
1UO d ≠ UO ôU ≠ WO D « U uKF*«Ë ÍuO(« ¡UB ù« r ≠ VD « WOK ≠ u uM ≈ WF U
1UO d ≠ UO ôU ≠ W u b « WO Ë_«Ë VKI « W «d r ≠ VD « WOK ≠ u uM ≈ WF U
2UO d ≠ UO ôU ≠ VKI « V r ≠ VD « WOK ≠ u uM ≈ WF U
3cemilcolak@yahoo.com
